CodeCompressionDomain AdaptationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: This paper proposes a method called SketchTune for compressing LLM weights through learned sketching and directly fine-tuning, balancing model compression and adaptation, eliminating the low-rank assumption and multi-path computation of traditional PEFT.
SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models
Wei Huang (University of Hong Kong), XIAOJUAN QI
CodeCompressionOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes SliM-LLM, a weight significance-driven grouped mixed-precision post-training quantization framework for compressing large language models.
π― What it does: A Social Co-evolution (SC) framework and DISC module are proposed, utilizing dynamic interactions between multi-task models to enhance existing task performance and learn new tasks.
Softmax is not Enough (for Sharp Size Generalisation)
Petar VeliΔkoviΔ (Google DeepMind), Razvan Pascanu (Google DeepMind)
CodeRetrievalOptimizationTransformerText
π― What it does: This paper studies the dispersion characteristics of softmax as the input size increases, proving that softmax must disperse with more input items, leading to an inability to maintain the robustness of 'sharp' functions (such as maximum value retrieval). It proposes an entropy-based adaptive temperature method to alleviate this issue.
π― What it does: A sequential Monte Carlo algorithm based on separated diffusion, DDSMC, has been developed to achieve asymptotically accurate posterior sampling in linear Gaussian Bayesian inverse problems using a pre-trained diffusion model as a prior.
Sorbet: A Neuromorphic Hardware-Compatible Transformer-Based Spiking Language Model
Kaiwen Tang (National University of Singapore), Weng-Fai Wong (National University of Singapore)
CodeComputational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerLarge Language ModelText
π― What it does: A Transformer-based spiking language model called Sorbet, fully compatible with brain-inspired hardware, has been designed and implemented. It achieves binary weights through knowledge distillation and quantization, reducing energy consumption.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelImageVideoText
π― What it does: A general, training-independent sparse attention mechanism called SpargeAttn is proposed, which can accelerate the inference of language, image, and video generation models without sacrificing model quality.
Rajiv Movva (University of California Berkeley), Emma Pierson (University of California Berkeley)
CodeExplainability and InterpretabilityComputational EfficiencyLarge Language ModelAuto EncoderText
π― What it does: Utilizing sparse autoencoders to learn interpretable hidden units, then selecting units related to the target variable through Lasso, and generating natural language hypotheses for these units using LLM, completing the generation of interpretable hypotheses regarding the relationship between text and the target variable.
π― What it does: A training-independent sparse video generation framework called Sparse VideoGen (SVG) is proposed, which accelerates the inference of video Diffusion Transformers by identifying and utilizing sparse patterns in spatial and temporal heads of 3D full attention.
Speculative Prefill: Turbocharging TTFT with Lightweight and Training-Free Token Importance Estimation
Jingyu Liu (University of Chicago), Ce Zhang (University of Chicago)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A training-free Speculative Prefill framework is proposed, which estimates and discards unimportant prompt tokens through a lightweight model, significantly reducing TTFT and improving maximum QPS during the prefill stage of LLM inference.
SPHINX: Structural Prediction using Hypergraph Inference Network
Iulia Duta (University of Cambridge), Pietro Lio
CodeGraph Neural NetworkPoint CloudGraph
π― What it does: This paper proposes an unsupervised hypergraph structure inference model called SPHINX, which can automatically learn implicit hypergraphs from point-level signals and be used for downstream tasks.
π― What it does: This paper proposes SpikeVideoFormer, an efficient temporal-driven Transformer capable of video classification, pose tracking, and semantic segmentation.
Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution
Jiayu Zhang (Suzhou University of Technology), Huaming Chen (University of Sydney)
CodeAnomaly DetectionExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
π― What it does: A novel OOD detection method S & I is proposed, which splits the intermediate layers of the network and integrates adversarial gradients along the reverse gradient path to obtain more reliable explanation patterns.
SPMC: Self-Purifying Federated Backdoor Defense via Margin Contribution
Wenwen He (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningImage
π― What it does: A self-purification federated learning backdoor defense method called SPMC is proposed, which measures the differences between clients using marginal contribution and dynamically weights aggregation, while achieving self-purification on the client side through gradient alignment.
π― What it does: This paper studies a novel unbiased similarity measurement metricβPure Square Euclidean Distance (PSED)βfor evaluating and optimizing class-level discriminability in deep networks.
Stable Fair Graph Representation Learning with Lipschitz Constraint
Qiang Chen (Central South University), Chang Xu (University of Sydney)
CodeOptimizationRepresentation LearningGraph Neural NetworkGraphTabularFinance Related
π― What it does: A stable fair graph representation learning framework (SFG) is proposed, aiming to maintain training stability while preserving accuracy and fairness.
Star Attention: Efficient LLM Inference over Long Sequences
Shantanu Acharya (NVIDIA), Boris Ginsburg (NVIDIA)
CodeRetrievalComputational EfficiencyTransformerLarge Language ModelTextSequentialBenchmark
π― What it does: A two-stage block sparse attention algorithm called Star Attention is proposed, which achieves efficient inference on long sequences through an anchor block mechanism, significantly reducing computational costs.
Statistical Hypothesis Testing for Auditing Robustness in Language Models
Paulius Rauba (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: A distribution-based perturbation analysis framework (DBPA) is proposed, which transforms the detection of output perturbations from large language models into a frequentist statistical hypothesis testing problem.
Stealing That Free Lunch: Exposing the Limits of Dyna-Style Reinforcement Learning
Brett Barkley (University of Texas at Austin), David Fridovich-Keil (University of Texas at Austin)
CodeReinforcement LearningTabularBenchmark
π― What it does: Evaluate the performance differences of Dyna-style model-based reinforcement learning on Gym and DMC benchmarks, and find that it generally performs worse than the non-Dyna baseline on DMC, revealing the no free lunch phenomenon.
Long-Kai Huang (Tencent AI Lab), Jianhua Yao (Tencent AI Lab)
CodeGenerationOptimizationDrug DiscoveryTransformerLarge Language ModelTextBiomedical Data
π― What it does: This paper proposes the migration of Activation Steering technology to Protein Language Models (PLM), utilizing fine-tuning of internal activations during inference to achieve control over the properties of protein sequence generation and optimization.
π― What it does: A framework called ShaRP is proposed and implemented, which solves imaging inverse problems by randomly sampling a set of pre-trained deep recovery models (not limited to denoisers) as priors.
Stochastic Encodings for Active Feature Acquisition
Alexander Luke Ian Norcliffe (University of Cambridge), Pietro Lio (University of Cambridge)
CodeClassificationAuto EncoderTabularBiomedical Data
π― What it does: A method for active feature acquisition based on a stochastic encoder, SEFA, is proposed to gradually select features during testing to improve prediction accuracy.
Stochastic Poisson Surface Reconstruction with One Solve using Geometric Gaussian Processes
Sidhanth Holalkere (Cornell University), Alexander Terenin (Cornell University)
CodeOptimizationGaussian SplattingPoint Cloud
π― What it does: This paper proposes a novel stochastic Poisson surface reconstruction method based on geometric Gaussian processes, which can simultaneously complete interpolation and implicit surface reconstruction with a single linear solve, thus achieving a reconstruction process that is sensitive to local queries and outputs.
SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics
Suyuan Zhao (Tsinghua University), Zaiqing Nie (Tsinghua University)
CodeSegmentationDomain AdaptationTransformerContrastive LearningBiomedical Data
π― What it does: A multi-scale spatial transcriptomics foundational model (SToFM) has been constructed, capable of simultaneously integrating macroscopic tissue structure, microscopic cell interactions, and gene expression information.
Strategic Planning: A Top-Down Approach to Option Generation
Max Ruiz Luyten (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a top-down reinforcement learning framework (Strategist) based on human strategic heuristics, which generates multi-layer strategy trees using LLM and transforms natural language strategies into quantifiable reward shaping, thereby accelerating the convergence of RL in sparse reward and exploration-constrained real-world environments.
Strong and Weak Identifiability of Optimization-based Causal Discovery in Non-linear Additive Noise Models
Mingjia Li (East China Normal University), Aimin Zhou
CodeOptimizationGraphTabularBiomedical Data
π― What it does: This paper addresses causal discovery under the nonlinear additive noise model (ANM) by categorizing the identifiability problem into strong identifiability and weak identifiability. Based on this, a unified optimization search method called GENE is designed, which evaluates causal order using both residual independence and goodness of fit, solving the issue of traditional methods performing poorly in weak identifiability scenarios.
Stronger Neyman Regret Guarantees for Adaptive Experimental Design
Georgy Noarov (University of Pennsylvania), Aaron Roth (Amazon Web Services)
CodeTabularTime SeriesFinance Related
π― What it does: This paper proposes two adaptive experimental design methods (ClipOGD SC and MGATE) that achieve low variance adaptive weight allocation for estimating average treatment effects in a finite sample setting.
Subgoal-Guided Policy Heuristic Search with Learned Subgoals
Jake Tuero (University of Alberta), Levi Lelis
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderTabular
π― What it does: This paper proposes a sub-goal based strategy-guided tree search method, which utilizes online learning of sub-goal generators and multi-level strategies from both successful and failed tree data during the search process, significantly improving the sample efficiency of the search.
Qingtian Zhu (University of Tokyo), Yinqiang Zheng (University of Tokyo)
CodeGraph Neural NetworkAuto EncoderBiomedical Data
π― What it does: This paper proposes a continuous sparse implicit neural representation named SUICA for ultra-high-dimensional sparse modeling and reconstruction of spatial transcriptomics data.
Sundial: A Family of Highly Capable Time Series Foundation Models
Yong Liu (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeTransformerFlow-based ModelTime SeriesFinance Related
π― What it does: A family of foundational models for time series, called Sundial, is proposed based on flow matching, which can perform unsupervised pre-training directly on a continuous value domain without discretization and supports diverse generative predictions.
π― What it does: Reformulates survival analysis as a density estimation problem and proposes a two-step algorithm: first, estimate the cumulative hazard function using any density estimation model, and then obtain the individual survival function through post-processing with a known copula.
SynEVO: A neuro-inspired spatiotemporal evolutional framework for cross-domain adaptation
Jiayue Liu (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeDomain AdaptationGraph Neural NetworkContrastive LearningTime Series
π― What it does: This paper proposes a neuroscience-based evolvable spatiotemporal network called SynEVO, which utilizes curriculum learning-style sample rearrangement, elastic public containers, task-agnostic feature extractors, and adaptive dynamic couplers to achieve continuous learning and knowledge transfer for cross-domain spatiotemporal prediction.
π― What it does: By sampling images on the ANN perception boundary and combining them with human experiments, the variMNIST dataset is constructed to study and implement the prediction and manipulation of individualized perceptual differences.
Synthetic Text Generation for Training Large Language Models via Gradient Matching
Dang Nguyen (University of California), Baharan Mirzasoleiman (University of California)
CodeGenerationData SynthesisOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A theoretically rigorous algorithm is proposed to generate readable, privacy-preserving synthetic text for fine-tuning large language models.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: This paper proposes the T1 framework, which first conducts chain-of-thought (CoT) pre-training with trial-and-error and self-checking on a large model, and then further enhances reasoning capabilities using reinforcement learning (RL);
CodeClassificationDomain AdaptationData-Centric LearningLarge Language ModelTabularBenchmarkFinance Related
π― What it does: TabFSBench is proposed - the first benchmark for feature shift in tabular data, systematically studying the impact of feature shift on model performance and evaluating multiple models.
TabICL: A Tabular Foundation Model for In-Context Learning on Large Data
Jingang QU, Marine Le Morvan (INRIA Saclay)
CodeClassificationTransformerTabular
π― What it does: A scalable table-based model TabICL is proposed, utilizing a two-stage column-row attention mechanism to achieve single forward inference for large-scale tables;
TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data
Elias Chaibub Neto (Sage Bionetworks)
CodeData SynthesisSafty and PrivacyTabular
π― What it does: A lightweight, model-free, and non-parametric synthetic tabular data generation method called TabSDS is proposed, which approximates the joint distribution with the original data through ranking and data shuffling.
Targeted Unlearning with Single Layer Unlearning Gradient
Zikui Cai (University of California Riverside), M. Salman Asif (University of California Riverside)
CodeComputational EfficiencyData-Centric LearningVision Language ModelDiffusion modelImageMultimodality
π― What it does: This paper proposes a Single-Layer Unlearning Gradient (SLUG) method that achieves target information unlearning in large-scale multimodal models using a single gradient update.
π― What it does: The Task-Aware Virtual Training (TAVT) algorithm is proposed in meta reinforcement learning, utilizing virtual tasks to enhance generalization capabilities for out-of-distribution (OOD) tasks.
TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness
Cheng Huang (Zhejiang University of Technology), Peter AG Watson
CodeDiffusion modelMultimodalityTime Series
π― What it does: A multi-modal diffusion model, TCP-Diffusion, is proposed to predict precipitation around the center of tropical cyclones at any location globally for the next 12 hours.
Henrik Schopmans (Karlsruhe Institute of Technology), Pascal Friederich (Karlsruhe Institute of Technology)
CodeGenerationOptimizationDrug DiscoveryFlow-based ModelSequentialBiomedical Data
π― What it does: Train a reversible flow network by first fitting the Boltzmann distribution at high temperatures using reverse KL divergence, and then recursively cooling down through importance sampling and fine-tuning with forward KL to achieve molecular sampling without mode collapse.
Temporal Query Network for Efficient Multivariate Time Series Forecasting
Shengsheng Lin (South China University of Technology), Weiwei Lin (South China University of Technology)
CodeTransformerTime Series
π― What it does: A novel Temporal Query (TQ) technique is proposed, and based on this, a lightweight TQNet is constructed for multivariate time series forecasting.
Test-Time Adaptation for Online Vision-Language Navigation with Feedback-based Reinforcement Learning
Sungjune Kim, Sangpil Kim (Korea University)
CodeLarge Language ModelReinforcement LearningMultimodality
π― What it does: Proposes the FEEDTTA framework, which utilizes binary feedback at termination as a reward to achieve online visual-language navigation adaptation during testing through REINFORCE reinforcement learning.
Test-time Adapted Reinforcement Learning with Action Entropy Regularization
Shoukai Xu (South China University of Technology), Peilin Zhao (Tencent AI Lab)
CodeReinforcement LearningTabularBenchmark
π― What it does: Test-Time Adapted Reinforcement Learning (TARL) is proposed based on offline reinforcement learning. It fine-tunes the parameters of normalization layers using observed states only during testing, through unsupervised entropy minimization and KL constraints, to adapt to distribution shifts in online environments.
π― What it does: The FOCAL framework is proposed, which transforms the input image into the most typical view during the inference phase by generating and ranking various transformations, utilizing the visual priors of foundational models (CLIP, Stable Diffusion) to achieve robustness against multiple visual transformations.
Linjing You (Institute of Automation, Chinese Academy of Sciences), Xiayuan Huang (Beijing Forestry University)
CodeDomain AdaptationImage
π― What it does: Proposes the Test-time Correlation Alignment (TCA) method, which constructs a pseudo-source correlation matrix using high-confidence test samples, achieving feature covariance alignment through linear transformation without updating model parameters, thus enabling adaptation during testing.
Jinwu Hu (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeTransformerLarge Language ModelTextBenchmarkAgriculture RelatedFinance Related
π― What it does: This paper proposes a Test-Time Learning (TTL) framework called TLM, which dynamically updates large language models (LLMs) during inference using only unlabeled test data, achieving self-supervised adaptation through input perplexity minimization.
Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback
Yafu Li (Shanghai AI Laboratory), Yu Cheng (Chinese University of Hong Kong)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: A method for real-time preference alignment of large language models during inference (Test-Time Preference Optimization, TPO) is proposed, which does not require updating model parameters.
TGDPO: Harnessing Token-Level Reward Guidance for Enhancing Direct Preference Optimization
Mingkang Zhu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes the introduction of token-level reward guidance in Direct Preference Optimization (DPO) to improve the alignment training of LLMs.
π― What it does: This paper proposes a cell instance segmentation method based on the four-color theorem, transforming instance segmentation into four types of semantic segmentation.
The Polynomial Stein Discrepancy for Assessing Moment Convergence
Narayan Srinivasan (Queensland University of Technology), Leah F South
Code
π― What it does: A polynomial Stein discrepancy (PSD) with linear time complexity is proposed to assess the moment convergence of Bayesian sampling and conduct power tests.
The Ripple Effect: On Unforeseen Complications of Backdoor Attacks
Rui Zhang (University of Electronic Science and Technology of China), Yang Zhang (CISPA Helmholtz Center for Information Security)
CodeClassificationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A quantitative study of backdoor attacks in pre-trained language models is conducted, assessing their unintended effects on non-target downstream tasks and proposing mitigation strategies.
The Sparse-Plus-Low-Rank Quasi-Newton Method for Entropic-Regularized Optimal Transport
Chenrui Wang (Shanghai University of Finance and Economics), Yixuan Qiu (Shanghai University of Finance and Economics)
CodeOptimizationImage
π― What it does: A sparse + low-rank approximation quasi-Newton method is proposed for solving large-scale entropy-regularized optimal transport problems.
Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport
Jayadev Naram (Chalmers University of Technology), Giuseppe Durisi (Chalmers University of Technology)
CodeDomain AdaptationImage
π― What it does: This paper derives a generalization upper bound for the partial domain adaptation (PDA) problem based on partial optimal transport and proposes a theoretically guided weight allocation scheme and the corresponding learning algorithm WARMPOT.
Three-Dimensional Trajectory Prediction with 3DMoTraj Dataset
Hao Zhou (Great Bay University), Fei Luo (Great Bay University)
CodeAutonomous DrivingOptimizationRecurrent Neural NetworkAuto EncoderPoint CloudTime Series
π― What it does: This study investigates 3D trajectory prediction, proposing the 3DMoTraj dataset and a prediction framework that decouples the three axes and refines associations.
CodeConvolutional Neural NetworkContrastive LearningTime Series
π― What it does: A hard-coded invariant convolution layer is proposed, achieving invariance to amplitude scaling, offset, and smoothing trends in time series modeling.
TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation
Daoyu Wang (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
CodeClassificationRepresentation LearningTransformerDiffusion modelContrastive LearningTime Series
π― What it does: This paper proposes a self-supervised time series pre-training framework called TimeDART, which can simultaneously learn global trends and local details.
π― What it does: A multivariate long-term time series forecasting model called TimePro based on Mamba has been constructed, utilizing variable and time-aware hyperstates to model the multi-delay relationships between variables.
TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation
Hyeongwon Jang (KAIST), Eunho Yang (KAIST)
CodeExplainability and InterpretabilityRecurrent Neural NetworkTime SeriesElectronic Health Records
π― What it does: An interpretable method for time series models, TIMING, is proposed, along with new evaluation metrics: Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP) to more fairly measure attribution quality.
π― What it does: A label completion method based on transfer learning, TLLC, is proposed to address the issue of label missing due to sparse annotations by workers in crowdsourcing scenarios.
π― What it does: The Token Coordinated Prompt Attention (TCPA) module is proposed, which allocates dedicated prompts for CLS and image tokens in visual prompts and achieves interaction through an attention mechanism.
Towards a Mechanistic Explanation of Diffusion Model Generalization
Matthew Niedoba (University of British Columbia), Frank Wood (University of British Columbia)
CodeRestorationGenerationDiffusion modelImage
π― What it does: Analyze the generalization mechanism of diffusion models and propose a patch-based denoiser (PSPC) that is untrained to approximate the output of network denoisers.
π― What it does: A new Reward-Guided Text Generation (RGTG) method is proposed, aimed at improving generation efficiency and reducing computational overhead during inference by calling the reward model only once.
π― What it does: This paper studies the robustness of the Influence Function in deep networks, pointing out that the sharpness of the validation risk leads to the failure of traditional methods, and proposes a new influence function estimation method based on Flat Validation Minima.
Training High Performance Spiking Neural Network by Temporal Model Calibration
Jiaqi Yan (Zhejiang University), Gang Pan (Zhejiang University)
CodeClassificationSpiking Neural NetworkImageText
π― What it does: Proposes the Temporal Model Calibration (TMC) method, which utilizes gradient re-scaling in the time dimension (confidence regularization + exponential Ξ» constraint) to enhance the diversity of temporal logit gradients, achieving direct training of SNNs for temporal calibration.
π― What it does: A smooth SchrΓΆdinger bridge method is proposed, replacing traditional Brownian motion with a smooth high-order autoregressive Gaussian process to address noise and path roughness issues in particle trajectory inference.
TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation
Jaeho Kim (Ulsan National Institute of Science and Technology), Seulki Lee (Ulsan National Institute of Science and Technology)
CodeDomain AdaptationTime SeriesElectrocardiogram
π― What it does: A pseudo-labeling method named TransPL is designed for unsupervised time series domain adaptation, primarily generating pseudo-labels through the state transition matrix of vector quantization codes.
TypyBench: Evaluating LLM Type Inference for Untyped Python Repositories
Honghua Dong (University of Toronto), Xujie Si (University of Toronto)
CodeAI Code AssistantTransformerLarge Language ModelTextBenchmark
π― What it does: This paper proposes the TYPYBENCH benchmark, designs two new metrics (TYPESIM and TYPECHECK), and evaluates the type inference capabilities of various large language models on 50 high-quality Python repositories.
UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models
Xin Xu (Hong Kong University of Science and Technology), Yang Wang (Hong Kong University of Science and Technology)
CodeLarge Language ModelTextBenchmarkPhysics Related
π― What it does: UGPhysics has been constructed as a benchmark for undergraduate-level physics reasoning, and the MARJ evaluation framework has been proposed.
Ultra Lowrate Image Compression with Semantic Residual Coding and Compression-aware Diffusion
Anle Ke (Nanjing University), Zhan Ma (Nanjing University)
CodeRestorationCompressionLarge Language ModelDiffusion modelImage
π― What it does: A novel extreme low bitrate image compression framework called ResULIC is proposed, which can achieve high-quality reconstruction under conditions of less than 0.005 bpp.
UltraTWD: Optimizing Ultrametric Trees for Tree-Wasserstein Distance
Fangchen Yu (Chinese University of Hong Kong), Qiang Sun (Mohamed bin Zayed University of Artificial Intelligence)
CodeRetrievalOptimizationText
π― What it does: This paper proposes UltraTWD, an unsupervised framework that can simultaneously optimize tree structure and edge weights, making the tree-Wasserstein distance closer to the 1-Wasserstein distance.
π― What it does: This paper proposes the UncertainSAM method for rapid and efficient quantification of prediction uncertainty in the Segment Anything Model (SAM), which includes Bayesian approximation and a lightweight USAM MLP estimator.
Uncertainty Quantification for LLM-Based Survey Simulations
Chengpiao Huang (Columbia University), Kaizheng Wang (Columbia University)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A questionnaire simulation method based on large language models (LLM) is proposed, which constructs confidence intervals using generated synthetic answers and achieves reliable inference of real population parameters by adaptively selecting the simulated sample size.
Uncertainty-Based Extensible Codebook for Discrete Federated Learning in Heterogeneous Data Silos
Tianyi Zhang (University of Minnesota), Dianbo Liu (National University of Singapore)
CodeFederated LearningImage
π― What it does: A federated learning framework based on uncertainty-extended codebooks (UEFL) is proposed, which improves model accuracy and reduces uncertainty by dynamically expanding discrete vector codebooks under multi-data silos.
Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models
Shuoyuan Wang (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
CodeClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImage
π― What it does: This study addresses the issue of confidence miscalibration that arises after prompt tuning in CLIP, proposing the Dynamic Outlier Regularization (DOR) method.
Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods
Yifan HAO, Tong Zhang (University of Illinois Urbana-Champaign)
CodeOptimizationLarge Language ModelSupervised Fine-TuningText
π― What it does: Through experimental and theoretical analysis, the phenomenon of overfitting in supervised fine-tuning was studied, demonstrating that model ensembling can simultaneously improve downstream task performance and alleviate forgetting of upstream tasks.
Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees
Xin Yu (Pennsylvania State University), Runze Li (Pennsylvania State University)
CodeFederated LearningImageText
π― What it does: This paper studies the trade-off between statistical accuracy and communication efficiency in Personalized Federated Learning (PFL), providing the optimal statistical convergence rate and communication complexity under a multi-task/global + local model framework.
π― What it does: The UNIFEWS framework is proposed, which unifies the item-wise sparsification of graph structures and model weights to significantly reduce the computational load of GNNs;
π― What it does: This paper proposes a unified Continuous Weak Features Learning (cWFL) theoretical framework and provides a generalization error analysis.
π― What it does: The UNIMC framework is proposed, which implements multi-category and multi-instance human and animal keypoint-guided image generation on the Diffusion Transformer using a unified keypoint encoder and temporal-aware keypoint modulator, and constructs the HAIG-2.9M large-scale keypoint dataset.
Unisolver: PDE-Conditional Transformers Towards Universal Neural PDE Solvers
Hang Zhou (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeTransformerLarge Language ModelTime SeriesPhysics Related
π― What it does: Developed Unisolver, a Transformer-based PDE conditioning model capable of unifying the handling of various types of partial differential equations.
Unisoma: A Unified Transformer-based Solver for Multi-Solid Systems
Shilong Tao (Peking University), Yunhuai Liu (Peking University)
CodeTransformerPhysics Related
π― What it does: The Unisoma model is proposed, which addresses the physical simulation of multi-solid systems using an explicitly modeled Transformer framework.
Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing
Zijie Qiu (Fudan University), Siqi Sun (Fudan University)
CodeProtein Structure PredictionTransformerBiomedical Data
π― What it does: This paper presents RankNovoβa deep learning framework based on list-based reordering to improve the accuracy of de novo protein sequencing.
π― What it does: A two-stage unsupervised multi-object segmentation framework called unMORE is proposed, which first learns a three-layer object center-boundary representation on ImageNet, and then automatically locates and segments multiple objects in a single image through a network-free inference module.
Variational Phylogenetic Inference with Products over Bipartitions
Evan Sidrow (Simon Fraser University), Lloyd T Elliott
CodeReinforcement LearningTabularBiomedical Data
π― What it does: This paper proposes a variational Bayesian inference method for ultrametric trees, VIPR, by modeling the distribution of pairwise coalescent times and using single-linkage clustering to map to tree structures, resulting in a differentiable variational distribution.
VCT: Training Consistency Models with Variational Noise Coupling
Gianluigi Silvestri (OnePlanet Research Center), Yuki Mitsufuji (Sony Group Corporation)
CodeGenerationFlow-based ModelAuto EncoderImage
π― What it does: This paper proposes a framework that introduces variational noise coupling in consistency training, utilizing an encoder to learn data-related noise distributions and performing joint training through KL regularization.
π― What it does: This paper presents GRIMOIRE, a two-stage SVG generation framework based on raster image supervision, capable of generating, completing, and editing vector graphics from natural language text.
VideoRoPE: What Makes for Good Video Rotary Position Embedding?
Xilin Wei (Fudan University), Dahua Lin (Shanghai AI Laboratory)
CodeRetrievalTransformerVision Language ModelVideo
π― What it does: A new video position encoding scheme, VideoRoPE, is designed and evaluated to address four key attributes of RoPE in videos: 3D structure, frequency allocation, spatial symmetry, and temporal index scaling.
Visual Abstraction: A Plug-and-Play Approach for Text-Visual Retrieval
Guofeng Ding (Sichuan University), Xi Peng (Sichuan University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelImageVideoTextChain-of-Thought
π― What it does: A scheme is proposed to abstract visual content into natural language descriptions during the testing phase and perform text retrieval (VISA), thereby enhancing text-visual retrieval performance.
Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models
Mingi Jung (Seoul National University), Sungroh Yoon (Seoul National University)
CodeRecognitionGenerationTransformerLarge Language ModelImageTextMultimodality
π― What it does: A training-free visual attention recalibration method called SPARC is proposed, aimed at improving the accuracy and recall of multimodal large language models when generating detailed image descriptions.
Huayu Chen (Tsinghua University), Jun Zhu (Tsinghua University)
CodeGenerationDiffusion modelImage
π― What it does: A Guidance-Free Training (GFT) method is proposed to eliminate the dependence on dual model guidance (CFG) in visual generative models, directly training a single model to achieve low-temperature sampling.
CodeRepresentation LearningDrug DiscoveryTransformerAuto EncoderContrastive LearningImageBiomedical Data
π― What it does: A framework based on self-supervised learning is proposed for pre-training large-scale cell microscopy images, model expansion, and selecting the best representations through intermediate layer probing, resulting in biologically meaningful representations.
π― What it does: This paper proposes a Pareto front learning framework PHN-HVVS based on Voronoi grid partitioning and genetic algorithms for high-dimensional multi-objective optimization, applied to collaborative federated learning.